Maybe the tools eventually manage to get so good that they don't cause big issues unless someone checks the code but currently I wouldn't trust their code with anything important. Hobby projects, concepts/demos, helper and that sort of stuff is fine.

The whole "write exact specs, let AI do it's stuff then check" is just not a good fit for human brain, IMO. Firstly, how many have managed to write an exact spec out of the gate like that? WHen you then need to revisit AI tends to rewrite massive parts of the code like that. Nice if you just spent days reviewing it. Reviewing is boring but with AI you have to be vigilant because it will do a lot of the mistakes that programmer would have caught and fixed automatically.

So, I don't know how much those better models will actually improve productivity. Price will be eye-watering, that much is clear. We are talking tens of thousands or more per year per developer. If one company runs a local model pretty much for the fixed price of the hardware and another pays more every year per developer than that hardware cost productivity difference needs to be massive. I'm not quite sold on that yet.

As a case in point, I gave Claude a specification from a component vendor for a specific service I need. That specification was quite specific in what needs to be done. I checked the output and Claude had decided not to do large parts of it without telling me. I had to specifically list every step and tell it to check if they had been done. So the entire thing had to be rewritten pretty much from scratch. And this was with Sonnet High.